AI agents
What is an agent, how is it different from a bot, how to build the first one and when does it pay off. materials: 31
Media
Copperhead designs boards in KiCad and commits by stages
The open agent breaks the development of the board into eight stages, each must leave the artifact on disk and go through the ERC/DRC of KiCad itself, otherwise there will be no commit. The gate scheme is transferred to any agent pipeline, where there is something to check the result.
Desert Ant Labs and local models: what they pay for speed
Desert Ant Labs announced itself with a post about fast models on the device, the discussion collected 459 votes. There are no measurements, so we figure it out ourselves: what local inference gives, how they pay for it, and where it pays off today.
i-have-adhd: skill so that the agent responds with the first line
Ten rules in one SKILL.md: action from above, no introductions, the list is no longer than five points. The form of the answer is corrected, the correctness is not, and it is better to throw out the requirement to name the deadlines right away.
A month with an AI agent instead of yourself: what happened
In August, I gave my feed to an AI agent and said that if he couldn’t handle it, I’d send him back to vibe code. A month has passed, I tell you what happened and what it cost.
Knowledge base
Analogues of Claude Code: terminal agent, editor with model and add-on in IDE
The three classes of tools behave differently, and one thing distinguishes them is who runs the commands and reads their output. The mechanics of each class, what they pay and how to test the choice on your task.
Autonomous AI agents: where they can handle it themselves and where a person is required
Levels of autonomy from a hint to an action on a white list, the mechanics of independent work and a list of actions that are not given to the agent under any settings.
What is an AI agent: how does it differ from a chatbot and from a script
Dissection of the agent device in parts: model, tools, loop and stopping condition. Where is the border between a chatbot and a script, how to assemble a minimal cycle and what to look at to understand whether it works or coincides.
Claude Code agents: how subagents save context
The subagent works in its own window and returns a paragraph instead of the entire draft output. Where is the description file, what to write in the header, when the agent saves context and when it just adds forwarding.
Claude Code CLI: commands and flags for every day
Terminal agent in the project directory: launch subcommands, slash commands, permissions, project memory and headless mode for scripts - without unnecessary theory.
Claude Code: terminal agent for working with code
The agent lives in the terminal next to the working copy: it searches for files by content, edits them, runs tests and shows diff. An analysis of the mechanics, permissions, context and where the tool is dangerous.
Ready-made AI agents: when to take a box and when to build your own
Four levels of readiness, a table of pre-pilot selection criteria and a cheap verification procedure: shadow mode, a basket of real examples, the cost of a closed task.
II-agent for 1C: how to connect the model with the accounting system
Three layers of binding: how 1C sends data out, how named operations for the model are made from it, and why recording always goes through an extension with an idempotency key.
AI agents for business: which tasks are closed and which are not
Task suitability filter of five attributes, application map by department with an honest list of what agents do not pull, economics of a unit of work and a pilot’s route with metrics by which decisions are made.
AI agents for programming: what they really speed up in code
How the agent works with access to files, tests and git, what tasks it completes faster than a human, where it creates more work than it saves, and how to check its edits.
How to create an AI agent: from task description to working version
Step-by-step assembly route: task contract, cutting into tools, access rights, prompt, loop with limits, set of checks and gradual transfer of autonomy. With loop code and check file format.
Local AI agent: running the model on your own hardware
What is really needed for an agent to work on its own machine: an engine, quantization, memory for the KV cache, an OpenAI-compatible endpoint and a fair price.
MCP: protocol for connecting tools to the model
One contract instead of an adapter for each client-system pair: the server describes tools, resources and prompts, any supporting client picks them up. Analysis of handshakes, transports, risks and debugging.
Configuring the AI agent: parameters, permissions, limiters, debugging
Seven configuration nodes, a table “symptom → what to edit” and a set of checks, after which the agent stops going in circles and doing irreversible things.
Obsidian as a knowledge base with which the agent works
The Obsidian repository is a regular folder with markdown files, so it can be read equally by a human, a script, and an AI agent. What to put in frontmatter, how to give the agent access, how to prevent him from littering the database with duplicates, and how to check that the knowledge is there at all.
OpenAI API: what the request consists of
Analysis of the request body and response fields: what determines the form of the result, where money on tokens leaks, why the response breaks off mid-sentence, and when repeating a request makes sense and when not.
Answers to the same type of customer questions without an operator
How it works: classify requests by intent, build a knowledge base from operators’ answers, and let the bot reply only with a retrieved passage and source link. It hands everything else to a person with the conversation history.
AI agent platform: what to check before transferring processes there
Analysis by layers: models, execution cycle, tool registry, state, rights, trace and evaluation. What to ask the supplier and what to check with your hands before the first combat launch.
Accepting applications from the site: from the form to the task of the person in charge
How it works: the form saves each request to its own database before forwarding it, a model extracts fields from the text, and a rules table handles routing. A silence monitor catches forms that stop working.
Examples of AI agents: what tasks are given to them and how each one works inside
Seven standard agents - support, analytics over the database, error duty officer, documents, resource, content, orchestrator. For each: input, tools, output, weak point and metric.
Incoming email: from message to draft task
The mailbox collects clients, invoices, notifications and mailings into one pile, and tasks in the tracker live without a link to the letter. Analysis of mechanics: deterministic rules, analysis scheme, draft card, idempotency by thread.
Development of AI agents: cycle, tools, memory, control
What the agent is physically assembled from: a call cycle, tool diagrams, three layers of memory and limiters, without which it goes into disarray.
AI Agent System: Orchestration, Role Separation, and Boundaries
When one agent is not enough: bundle topologies, contracts between roles, shared memory, budgets and a trace that shows where the chain is broken.
Skill Claude Code: how the skill works and when to write it
Skill is a folder with the SKILL.md file, which is loaded only for a suitable task. What to write in the header, how the description differs from the body, and why design skills live in the repository.
Creation of AI agents: constructor, framework or own cycle
Three layers of agent assembly and what each gives and takes away: a visual designer, a code framework, a custom loop on the SDK. Selection criteria, assembly routes for each layer and rules for transition between them.
Vector search: why it is needed and when it is not needed
Embeddings, document slicing, ANN indexes and a hybrid with full-text search - the whole mechanics. Separately: a list of tasks where a vector is inferior to a regular word search, and how to measure this on your corpus.
Introducing AI agents: how to introduce an agent into a live process and not break it
The rights ladder from read to autonomy, shadow run, rollback point and metrics that show that the agent is actually helping the process.